A monitoring method for fracturing equipment fault classification model
By conducting multi-dimensional monitoring of the fault classification model for fracturing equipment, the problem of quality degradation during model deployment was solved, ensuring the stable operation and accurate diagnosis of the model.
Patent Information
- Application Number
- CN202410588875.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-13
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2044-05-13
AI Technical Summary
In the existing technology, the quality of the fault classification model of fracturing equipment may rapidly deteriorate during the production deployment process, resulting in inaccurate model output. Furthermore, there is a lack of effective monitoring methods to prevent model degradation, which may cause irreparable losses.
By using feature extraction, decision tree algorithm training, monitoring the variance and kurtosis of flow data, model performance indicators, and operational indicators, multi-dimensional monitoring of the fault classification model of fracturing equipment is achieved. This includes real-time detection of feature drift, model performance, and operational status, triggering alarms, and taking appropriate maintenance measures.
It has achieved continuous and stable operation of the fault classification model for fracturing equipment, and timely output accurate fault diagnosis results, thus avoiding losses caused by the degradation of model performance.
Smart Images

Figure CN118484649B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a monitoring method for a fault classification model of fracturing equipment. Background Art
[0002] Currently, many products use models to implement service scenarios and provide users with corresponding services. For example, the model is used to analyze the inlet flow data of the fracturing pump, and through the establishment of a fault diagnosis model for key components of the fracturing unit system (mainly the fracturing pump cylinder), automatic fault diagnosis is achieved, so that targeted repairs or replacements can be carried out, improving equipment efficiency and saving maintenance costs.
[0003] However, during the process of production deployment, the model may degrade without any warning and quickly reduce its quality. If appropriate maintenance measures are not taken in time to restore the performance of the model, it will not only affect the output of the model, but also cause irreparable losses to users.
[0004] In order to prevent the model from causing destructive effects, how to continuously monitor the performance of the model and ensure the accuracy of the prediction results is an urgent problem to be solved. Summary of the invention
[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for monitoring the fault classification model of fracturing equipment from four dimensions: input data drift, feature offset, model performance and model operation index. When a problem occurs in the model, appropriate maintenance measures can be taken in time to restore the model performance.
[0006] The object of the present invention is achieved through the following technical solution: A method for monitoring a fracturing equipment fault classification model comprises the following steps:
[0007] S1. Feature extraction: The waveform signal of the flow data of the key components of the fracturing unit system is input to extract features in the time domain and frequency domain; in the process of model establishment, the time domain features composed of the mean, variance, skewness, kurtosis, root mean square value, mean absolute deviation, peak value, peak factor, and range of the flow waveform, as well as the frequency domain features composed of the low-frequency sum, the middle-frequency sum, the high-frequency sum, the frequency corresponding to the maximum value of the spectrum kurtosis, and the frequency corresponding to the maximum value of the power spectrum are used as input data;
[0008] S2. Training fault classification model: Use decision tree algorithm to sort 14 features, quantify feature importance by recording total number of splits and total / average information gain of each feature, sort the number of splits or total / average information gain from high to low, check the feature sorting results and select the top five features as input features of the classification model;
[0009] Use the decision tree algorithm for classification modeling. The decision tree algorithm selects the optimal feature partition node by calculating information gain, creates child nodes and divides data according to feature values, uses pruning to avoid overfitting, classifies samples into leaf nodes according to the tree path, and outputs the final fault classification result;
[0010] S3, monitoring the variance and kurtosis of the flow data, when the difference between the variance of the real-time flow data and the variance of the baseline data exceeds a preset threshold, or the difference between the kurtosis and the sum of the kurtosis of the baseline data exceeds a preset threshold, an alarm is triggered, and the flow data of the key components of the fracturing unit system are re-input, and the process returns to step S1 to re-perform feature extraction and model training;
[0011] S4. If the difference between the variance and kurtosis of the real-time traffic data and the variance and kurtosis of the baseline data does not exceed the preset threshold, the various indicators of the current fault classification model are explored; the accuracy and F1 value of the fault classification model are monitored and the threshold is set. The accuracy indicates the proportion of correctly classified samples to the total number of samples, and the F1 value is used to measure the performance of the model in maintaining a balance between precision and recall; when the accuracy and F1 value of the fault classification model are both lower than the threshold, it means that the current model is not competent for the fault classification task, triggering an alarm, adjusting the parameters of the fault classification model, and retraining the model;
[0012] S5. When the accuracy and F1 value of the fault classification model meet the standards, explore each feature of the data; sort the number of splits or total / average information gain of the 14 features and compare them with the sorting results of the original data; when the sorting results of several selected features change, it means that the currently selected features are no longer the most important features and cannot be used as the basis for the model to judge the fault type, triggering an alarm; reselect the top five features as the input features of the model and retrain the model;
[0013] S6. Monitor various indicators of the model running service, take the model's CPU, GPU and memory usage, and the corresponding response time of the current model running service as running indicators, and set thresholds for various running indicators; when any running indicator of the model exceeds the set threshold, it indicates that the current server has a problem, triggering an alarm, and the server is processed through operation and maintenance means.
[0014] The beneficial effects of the present invention are: monitoring the fracturing equipment fault classification model from four dimensions: input data drift, feature offset, model performance, and model operation index; when a problem occurs in the model, appropriate maintenance measures can be taken in time to restore the model performance; enabling the model to operate continuously and stably, maintain a stable prediction effect, and output fault diagnosis results in time. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 The present invention is a flow chart of a method for monitoring a fracturing equipment fault classification model. DETAILED DESCRIPTION
[0016] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.
[0017] like Figure 1 As shown, a fracturing equipment fault classification model monitoring method of the present invention comprises the following steps:
[0018] S1. Feature extraction: The waveform signal of the flow data of the key components of the fracturing unit system is input to extract features in the time domain and frequency domain. In the process of model establishment, the time domain features composed of the mean, variance, skewness, kurtosis, root mean square value, mean absolute deviation, peak value, peak factor, and range of the flow waveform are used, as well as the frequency domain features composed of the low frequency sum, the middle frequency sum, the high frequency sum, the frequency corresponding to the maximum value of the spectrum kurtosis, and the frequency corresponding to the maximum value of the power spectrum. A total of 14 features are used as input data, recorded as X1, X2, X3, ..., X 14 ;
[0019] S2. Training fault classification model: Use decision tree algorithm to sort 14 features, quantify feature importance by recording total number of splits and total / average information gain of each feature, sort the number of splits or total / average information gain from high to low, check feature sorting results and select the top five features (X1, X2, X3, X4, X5) as input features of the classification model. Generally, the top five features are mean absolute deviation, range, frequency corresponding to maximum power spectrum amplitude, kurtosis, and root mean square value.
[0020] Use the decision tree algorithm for classification modeling. The decision tree algorithm selects the optimal feature partition node by calculating information gain, creates child nodes and divides data according to feature values, uses pruning to avoid overfitting, classifies samples into leaf nodes according to the tree path, and outputs the final fault classification result;
[0021] By extracting the features of the waveform signals of the flow data of the key components of the fracturing unit system in the time and frequency domains, classification modeling is performed according to the three types of fracturing pump faults: suction port leakage, suction port blockage, suction port leakage and blockage, and normal state. A fault diagnosis model for the key components of the fracturing unit system (mainly the fracturing pump cylinder) is established to achieve automatic fault diagnosis.
[0022] S3. Input drift monitoring: In actual application, the various indicators of the data often change due to aging of data acquisition equipment, changes in environmental factors, and changes in industry and mining, which causes the distribution of the collected flow data to shift, resulting in a decrease in the accuracy of the model for fault diagnosis. Therefore, the variance and kurtosis of the flow data are monitored. The variance measures the deviation from the mean. Its increase indicates that the discreteness of the data increases and the reliability decreases; the kurtosis measures the extreme situation. When the kurtosis increases, it indicates that the data is sharper and there may be more extreme values. When the difference between the variance of the real-time flow data and the variance of the baseline data exceeds the preset threshold, or the difference between the kurtosis and the sum of the kurtosis of the baseline data exceeds the preset threshold, an alarm is triggered, and the flow data of the key components of the fracturing unit system is re-entered, and the step S1 is returned to re-feature extraction and model training to obtain a reliable model that adapts to the new data distribution.
[0023] S4. Model performance monitoring: If the difference between the variance and kurtosis of the real-time traffic data and the variance and kurtosis of the baseline data does not exceed the preset threshold, the various indicators of the current fault classification model are explored; the accuracy and F1 value of the fault classification model are monitored and thresholds are set. The accuracy indicates the proportion of correctly classified samples to the total number of samples, and the F1 value is used to measure the performance of the model in maintaining a balance between precision and recall; when the accuracy and F1 value of the fault classification model are both lower than the threshold, it means that the current model is not competent for the fault classification task, triggering an alarm, adjusting the parameters of the fault classification model, and retraining the model to achieve higher accuracy and F1 values.
[0024] S5. Feature drift monitoring: When the accuracy and F1 value of the fault classification model meet the standards, explore each feature of the data; sort the number of splits or total / average information gain of the 14 features and compare them with the sorting results of the original data; when the sorting results of several selected features change, it means that the currently selected features (mean absolute deviation, range, frequency corresponding to the maximum amplitude of the power spectrum, kurtosis, root mean square value) are no longer the most important features and cannot be used as the basis for the model to judge the fault type, triggering an alarm; reselect the top five features as the input features of the model and retrain the model;
[0025] S6. Monitor various indicators of the model running service, take the model's CPU, GPU and memory usage, and the corresponding response time of the current model running service as running indicators, and set thresholds for various running indicators; when any running indicator of the model exceeds the set threshold, it indicates that the current server has a problem, triggering an alarm, and processing the server through operation and maintenance means, so that the model can quickly respond to real-time data and quickly output fault diagnosis results.
[0026] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.
Claims
1. A monitoring method for a fracturing equipment fault classification model, characterized in that: The following steps are involved: S1. Feature extraction: The waveform signal of the flow data of the key components of the fracturing unit system is input to extract features in the time domain and frequency domain; in the process of model establishment, the time domain features composed of the mean, variance, skewness, kurtosis, root mean square value, mean absolute deviation, peak value, peak factor, and range of the flow waveform, as well as the frequency domain features composed of the low-frequency sum, the middle-frequency sum, the high-frequency sum, the frequency corresponding to the maximum value of the spectrum kurtosis, and the frequency corresponding to the maximum value of the power spectrum are used as input data; S2. Training fault classification model: Use decision tree algorithm to sort 14 features, quantify feature importance by recording total number of splits and total / average information gain of each feature, sort the number of splits or total / average information gain from high to low, check the feature sorting results and select the top five features as input features of the classification model; Use the decision tree algorithm for classification modeling. The decision tree algorithm selects the optimal feature partition node by calculating information gain, creates child nodes and divides data according to feature values, uses pruning to avoid overfitting, classifies samples into leaf nodes according to the tree path, and outputs the final fault classification result; S3, monitoring the variance and kurtosis of the flow data, when the difference between the variance of the real-time flow data and the variance of the baseline data exceeds a preset threshold, or the difference between the kurtosis and the sum of the kurtosis of the baseline data exceeds a preset threshold, an alarm is triggered, and the flow data of the key components of the fracturing unit system are re-input, and the process returns to step S1 to re-perform feature extraction and model training; S4. If the difference between the variance and kurtosis of the real-time traffic data and the variance and kurtosis of the baseline data does not exceed the preset threshold, various indicators of the current fault classification model are explored; Monitor the accuracy and F1 value of the fault classification model and set thresholds. The accuracy indicates the proportion of correctly classified samples to the total number of samples, and the F1 value is used to measure the performance of the model in balancing the precision and recall. When both the accuracy and F1 value of the fault classification model are lower than the threshold, it means that the current model is not competent for the fault classification task, triggering an alarm, adjusting the parameters of the fault classification model, and retraining the model. S5. When the accuracy and F1 value of the fault classification model meet the standards, explore each feature of the data; sort the number of splits or total / average information gain of the 14 features and compare them with the sorting results of the original data; when the sorting results of several selected features change, it means that the currently selected features are no longer the most important features and cannot be used as the basis for the model to judge the fault type, triggering an alarm; reselect the top five features as the input features of the model and retrain the model; S6. Monitor various indicators of the model running service, take the model's CPU, GPU and memory usage, and the corresponding response time of the current model running service as running indicators, and set thresholds for various running indicators; when any running indicator of the model exceeds the set threshold, it indicates that the current server has a problem, triggering an alarm, and the server is processed through operation and maintenance means.
